The ubiquity of social media has transformed online interactions among\nindividuals. Despite positive effects, it has also allowed anti-social elements\nto unite in alternative social media environments (eg. Gab.com) like never\nbefore. Detecting such hateful speech using automated techniques can allow\nsocial media platforms to moderate their content and prevent nefarious\nactivities like hate speech propagation. In this work, we propose a weak\nsupervision deep learning model that - (i) quantitatively uncover hateful users\nand (ii) present a novel qualitative analysis to uncover indirect hateful\nconversations. This model scores content on the interaction level, rather than\nthe post or user level, and allows for characterization of users who most\nfrequently participate in hateful conversations. We evaluate our model on 19.2M\nposts and show that our weak supervision model outperforms the baseline models\nin identifying indirect hateful interactions. We also analyze a multilayer\nnetwork, constructed from two types of user interactions in Gab(quote and\nreply) and interaction scores from the weak supervision model as edge weights,\nto predict hateful users. We utilize the multilayer network embedding methods\nto generate features for the prediction task and we show that considering user\ncontext from multiple networks help achieving better predictions of hateful\nusers in Gab. We receive up to 7% performance gain compared to single layer or\nhomogeneous network embedding models.\n